#TTS Playground (Voice Cloning)

TTS Playground is the dedicated testing interface for speech synthesis models (such as Qwen3-TTS) in Herdsman. This guide focuses on the Voice Clone feature: provide a reference audio clip and have the model speak new text in the cloned voice.

#Launching a TTS Model

1. Open the Models page.

2. Locate a model tagged TTS or Speech Synthesis (e.g., Qwen3-TTS Voice Clone).

3. Click Launch Now on the model card.

4. Wait for the status to switch to Ready, then click the card to open the Playground.

   

#Interface Overview

The TTS Playground is divided into a parameter panel (left), a script workspace (center), and a runtime monitor (right).

#Parameters (left sidebar)

• Basic settings:
  • Language — dropdown (e.g., Auto) selecting the language of the synthesis text.
  • Speed — slider (default 1). Drag right to speed up, left to slow down.
• Advanced settings:
  • Output mode:
    • Non-streaming (default) — waits until the full audio is generated, then plays it. Best for short texts.
    • Streaming / Realtime — plays as audio is generated. Lower latency (subject to model support).
• Voice (Clone core):
  • Reference Audio — the voice you want to imitate. Accepts a local file path, a URL, or Base64-encoded data.
  • Reference Text — the exact transcript of the reference audio. Providing an accurate transcript significantly improves cloning accuracy.

#Script Workspace (center)

• Text input:
  • Enter the text you want the model to speak in the top text box.
  • Example: "Hello — how is your day going? Welcome to Herdsman voice playground."
  • Supports mixed Chinese-English text and punctuation-driven pauses.
• Tags — the top-right corner provides Chinese, English, and Emotion tags to quickly switch input hints or adjust style.
• Output area:
  • Caption / waveform — generated captions appear here.
  • Audio player — standard controls (play/pause, scrub, volume) at the bottom.
• Action buttons:
  • Generate — start synthesis.
  • Stop — cancel the current job.

#Runtime Monitor (right sidebar)

• Runtime status:
  • TTFA / RTF — time-to-first-audio and real-time factor.
  • Latency / chunks — generation latency and chunk size.
  • Hardware usage — live CPU, GPU, and memory monitoring.
• Run log:
  • Shows model load events (e.g., Selected qwen3-tts-voiceclone) and audio initialization details.

#Voice Cloning Workflow

1. Prepare assets:
   • A clean reference audio clip (5–10 seconds is ideal).
   • The exact reference text spoken in the clip.
2. Configure cloning:
   • Enter the audio path or URL in Reference Audio.
   • Enter the transcript in Reference Text.
3. Write your script — type the text you want the model to read in the center workspace.
4. Adjust parameters — change Speed if needed (e.g., 1.2 for slightly faster delivery).
5. Generate and review:
   • Click Generate at the bottom left.
   • Use the audio player to listen. If the voice doesn't sound right, check that the reference audio is clean and verify that the reference text matches it exactly.

---

Notes:

• Reference text matters. For voice cloning, the reference text must accurately match the reference audio — mismatches degrade cloning quality.
• Audio format support. Reference audio typically supports WAV and MP3, depending on the model backend.